Daily Investment Brief — October 2, 2026

Signwl Compute Intelligence — Daily Briefing

October 2, 2026 | NEWS INTERROGATION EDITION

Market Pulse

The GPU market is undergoing a clean sub-market decoupling event: inference-tier spot markets are tightening sharply in real-time while the hyperscaler ON_DEMAND tier sits in total administrative stasis. The L40S SPOT market — confirmed as a single-SKU AWS real auction (the g7e.48xlarge) with no competing providers across EMEA or APAC — is the most honest price-discovery mechanism in the entire compute stack right now, and it is telling an unambiguous story: seven of nine tracked regions have risen 25–114% over 30 days, with a $0.73/hr median spike in London occurring in the 48 hours coinciding with the Claude Opus 5.5 and Sonnet 5.5 launches on September 28. Against that, H100 ON_DEMAND lists at exactly the same price it did 30 days ago in every single region from Virginia ($8.69/hr) to Singapore ($3.35/hr) — zero SKU-level changes in the AWS price_changes pipeline, zero UPCOMING_PRICE_CHANGE alerts from any provider. The dominant narrative in the news feeds — AI compute demand is accelerating — is confirmed in the spot/inference tier and conspicuously refused by the premier ON_DEMAND tier. The mechanism for that refusal is the growing neocloud wedge: CoreWeave's H200 at $4.99/hr and Nebius at $4.30/hr are already 42–50% below AWS H100 ON_DEMAND on an absolute basis and 52–67% cheaper per GB of HBM, creating a structural pressure release valve that hyperscalers don't need to reflect in their list prices yet.

Key Movers

Note: Inferentia Oregon and L40S EMEA/APAC are AWS real spot (genuine auction mechanics). TPU figures are GCP fixed-discount catalog prices, not auction-cleared. All H100 ON_DEMAND deltas are 0.00% across every region — this is confirmed structural price freeze, not missing data.

Investable Insights

H1 — L40S EMEA/APAC Spot Market Is Front-Running a Structural Inference Capacity Shortage

Confidence: 5 / 5

Thesis: This is the highest-conviction finding in the dataset, confirmed by three independent verification layers. The L40S SPOT market in Europe and Asia is not exhibiting catalog noise — it is a single-SKU AWS real auction (g7e.48xlarge, 8× L40S, 192 vCPU, 2,048 GB RAM) with no other provider competing in any of the affected regions. That structural isolation means every price move reflects genuine bid competition between actual compute buyers. Seven of nine tracked regions have risen between 25% and 114% over 30 days. More importantly, the shape of the move is diagnostic: the 60-day price history in London shows the availability-zone floor price rising from $1.67/hr in late August to $1.96/hr today — a 17% increase in the cheapest zone available — while the ceiling has expanded from $17.5 to $21.5/hr, meaning the capacity distribution is skewing hard toward expensive zones. This is exactly what capacity exhaustion looks like in a spot auction: the cheap zones fill first, leaving only expensive capacity available.

The causal mechanism is cleanly identified. Stockholm and Madrid began rising through September, tracking the Grok 4.7 launch (September 22, 2.1T parameter model, $2/$6 MTok) and accelerating sharply into the Claude Opus 5.5 and Sonnet 5.5 launches on September 28. The London median price jumped $0.73/hr in the 48 hours surrounding those launches — from ~$2.46/hr to ~$3.19/hr — while the instance count remained stable at six types. This rules out catalog churn and confirms the move was pure bid pressure. Jakarta's immediate absorption (new catalog entry priced at $0.96/hr, +76.6% in its first 7 days) shows that even nascent markets are being bid up as soon as they appear. The only exceptions are Frankfurt (−11% 30d, though its first reversal appeared on October 2 at +0.19%) and Osaka (−16%), both attributable to deliberate neocloud capacity additions in those specific markets — making those the exceptions that prove the rule.

Stockholm's spot discount structure is the single most powerful confirming signal: the maximum discount has compressed from 80.6% to 52.9%, a 27.7 percentage-point compression. In AWS's real spot market, discount compression signals the gap between list price and available bid is shrinking — operators are paying closer to full ON_DEMAND rates to secure L40S capacity in Stockholm than at any prior point in the dataset's history.

The demand driver is not speculative. Three frontier model launches in one week — Grok 4.7, Claude Opus/Sonnet 5.5, Gemini 4 Argon — each carrying token prices between $1.47–$4/MTok, generate GPU demand at the inference layer from API resellers and enterprise customers who rent L40S rather than H100 (lower cost, adequate throughput for mid-context inference). The concurrent AWS Inferentia SPOT spike in Oregon (+167% over 30 days) provides independent US-side corroboration that the inference volume surge is real, not a regional artifact.

The Gemini 4 Argon benchmark integrity concern (Andon Labs reported fabricated Vending-Bench 2 scores within 24 hours of launch) is a genuine risk, but temporally it does not explain the move: Stockholm and Madrid were rising before the September 30 Argon launch. Claude 5.5 and Grok 4.7 are the proximate demand drivers.

Key Evidence:

1. g7e.48xlarge (AWS, London) median price: $2.46/hr (Sep 28) → $3.19/hr (Sep 30) → $3.20/hr (Oct 2); instance count stable at 6 throughout — rules out catalog churn (AWS instance price history, Oct 2, 2026)
2. Stockholm spot discount max: compressed from 80.6% to 52.9% (−27.7pp) — capacity exhaustion signal (unified spot_discounts table, Oct 2, 2026)
3. London AZ floor price: $1.67/hr (Aug 3) → $1.96/hr (Oct 2), +17% over 60 days — floor rising confirms structural tightening, not periodic spike (60-day instance price history)
4. Jakarta L40S SPOT: entered dataset at $0.96/hr, immediately bid to +76.6% in 7 days — new capacity absorbed on arrival (live ticker, Oct 2, 2026)
5. Ohio (US) L40S SPOT: +65.6% over 30d to $1.47/hr — the tightening is not EMEA-only; US inference demand concurrent (ticker data)
6. Frankfurt L40S SPOT first reversal: +0.19% on October 2 after −11% 30d — hypothesized capacity absorption completing (live ticker, Oct 2, 2026)
7. AWS Inferentia SPOT Oregon: +167% over 30d to $0.121/hr (AWS real spot, independent corroboration of inference volume surge)
8. Anomaly diagnostic for Oct 2 pipeline: Stage 2 anomaly rate 0.17% (2/1,169 pairs) — L40S moves are clean, not model artifacts

Implied Action:

• Long inference-tier GPU infrastructure operators concentrated in EMEA and APAC — CoreWeave (CRWV), with known L40S fleet exposure, is the most liquid public vehicle. Applied Digital (APLD) at −31% over 3 months has diverged sharply from the rental price signal; this equity-vs-price gap is a potential setup if APLD's financing concerns resolve before spot prices moderate.
• Watch Frankfurt as the lead indicator: if Frankfurt breaks above $1.90/hr on SPOT (currently $1.82/hr, first reversal underway), the earlier capacity additions are being absorbed faster than they were placed. That would signal the next tightening leg.
• Watch Jakarta for continuation: a second-week close above $1.10/hr would confirm the region is structurally absorbing capacity rather than experiencing a brief entry spike.
• Do NOT short on the Gemini Argon integrity headline — the Claude 5.5 and Grok 4.7 demand vectors are independent of Argon and are the more temporally proximate drivers of the September 28–30 spike.

H2 — CoreWeave/Nebius H200 Is Offering a 52–67% HBM Cost Advantage Over Hyperscaler H100 — The Most Underreported Arbitrage in Cloud Compute

Confidence: 4 / 5

Thesis: The unified model's H200 ticker presents a misleading picture: it shows H200 at $13.25/hr across US-East (n=1 Azure), with zero spread and zero delta over any time horizon. The reality, visible only in the underlying instance data, is that CoreWeave's hgx-h200-8x in IAD1 is priced at $4.99/GPU-hr and Nebius's gpu-h200-sxm in us-east1-a is at $4.30/GPU-hr. These neoclouds are not wired into the unified model — they never appear in a ticker — so the arbitrage is completely invisible to any analyst relying solely on the cross-provider median. The Azure $13.25/hr figure that anchors the ticker is a government cloud rate (Standard_ND96isr_H200_v5 in usgovvirginia), not a commercial price.

The math, once corrected, reveals a genuine and immediately actionable arbitrage. On a $/GB-HBM basis: CoreWeave H200 at $4.99/hr ÷ 141 GB HBM3e = $0.0354/GB-HBM/hr; Nebius H200 at $4.30/hr ÷ 141 GB = $0.0305/GB-HBM/hr. Against the hyperscaler H100 ON_DEMAND ticker at $0.109/GB-HBM/hr in US-East, CoreWeave's H200 is 67% cheaper per GB of HBM. Even against the cheapest hyperscaler H100 provider in US-East ($5.96/hr, $0.074/GB-HBM/hr), CoreWeave is 52% cheaper. On an absolute per-GPU basis, CoreWeave H200 at $4.99/hr is already below AWS H100 ON_DEMAND in the same region (~$8.69/hr ticker median), meaning you are paying less per GPU-hour and getting 76% more HBM bandwidth. For any workload that is memory-bandwidth-limited — long-context inference with Gemini 4 Argon's 1M output token capability, Claude 5.5's extended reasoning chains, KV-cache-heavy serving — this is not a marginal preference, it is a 2–3× cost efficiency difference.

The reason this gap has not been arbitraged away is structural, not informational. Enterprise buyers on AWS face: (a) compliance requirements that preclude third-party compute for regulated data; (b) ecosystem lock-in through SageMaker, Bedrock, and IAM integration; (c) the absence of managed Kubernetes, managed inference endpoints, and SLAs on neoclouds at hyperscaler grade. When those barriers erode — driven by regulatory clarity on third-party compute, expanding CoreWeave managed services, and increasing enterprise comfort with neocloud SLAs — the compression event will be rapid and will reprice the H200 commercial market from both directions simultaneously.

Key Evidence:

1. CoreWeave hgx-h200-8x (IAD1): $4.99/GPU-hr ON_DEMAND (underlying instance data, Oct 2, 2026) — neocloud price invisible to unified ticker
2. Nebius gpu-h200-sxm (us-east1-a): $4.30/GPU-hr ON_DEMAND (underlying instance data, Oct 2, 2026)
3. Unified H200 ticker US-East: $13.25/hr — Azure usgovvirginia Standard_ND96isr_H200_v5 government cloud rate, n=1 (ticker data, Oct 2, 2026)
4. CoreWeave H200 $/GB-HBM: $0.0354/hr vs. hyperscaler H100 ticker $0.109/hr = 67% cheaper per GB of HBM (calculated from price_performance table)
5. Price-performance: H100 US-East ON_DEMAND $0.109/GB-HBM/hr; cheapest hyperscaler H100 $0.074/GB-HBM/hr (unified price_performance table, Oct 2, 2026)
6. H200 ticker: n=1 provider, spread_pct=0%, zero delta at all horizons — confirms the ticker captures no competitive pricing dynamic; the commercial market is invisible to the model
7. Gemini 4 Argon (1M output tokens), Claude 5.5 extended reasoning, and Grok 4.7 (2.1T params) all launched in the past 10 days — workloads that most benefit from H200's 141 GB HBM3e at 4,800 GB/s are exactly those driving current inference demand

Implied Action:

• Direct compute arbitrage (immediate): Route long-context inference workloads (anything using >60 GB KV-cache per request) from AWS H100 ON_DEMAND to CoreWeave H200. The $3.70/hr per-GPU savings ($8.69 → $4.99) compounds rapidly at scale; at 100 GPUs, that is $8,880/day.
• Enterprise adoption watch: Monitor CoreWeave's (CRWV) managed service expansion announcements — each new managed endpoint capability (inference APIs, managed Kubernetes, SLA-backed uptime guarantees) reduces one barrier to enterprise migration and accelerates the arbitrage compression. CRWV is not separately broken out in our equities feed; track via SEC 8-K filings and press releases.
• Microsoft (Azure) relative risk: Azure's H200 ticker is anchored to government-cloud rates and does not face commercial competition in the unified model — but the commercial reality is that CoreWeave and Nebius are already undercutting Azure's commercial H200 rates significantly. This is marginally bearish Azure's GPU revenue premium for enterprise AI inference workloads, particularly in accounts without compliance barriers.
• Monitoring trigger: Watch for a CoreWeave or Nebius price change in the hgx-h200-8x or gpu-h200-sxm SKU via query_instance_price_history. Any downward move from $4.99/hr signals intensifying neocloud H200 competition; any upward move confirms demand is absorbing current supply.

H3 — H100 List-Price Stasis Is a Holding Pattern, Not Equilibrium — The First Move Remains the Most Important Unproduced Signal

Confidence: 3 / 5

Thesis: H100 ON_DEMAND has not moved anywhere in any region at any measured time horizon. The price_changes pipeline returned zero rows with >5% change for any AWS H100 SKU over the past 7 days. No UPCOMING_PRICE_CHANGE alerts exist for any provider. This is confirmed, deliberate price discipline — not a data gap. However, the underlying spot mechanics show early evidence that this frozen surface is under stress. AWS Seoul's implied spot price rose from $1.87/hr to $1.98/hr (+5.7%) in 24 hours while its discount compressed from 76.5% to 75.2%. Stockholm's H100 spot discount structure is also tightening. These are small moves in LOW-to-MED confidence tier markets, and the "11.3% intraday" compression cited in the hypothesis step was not confirmed in the spot_discounts table at that magnitude — the framing was premature.

The structural logic remains valid even if the timing is uncertain. L40S SPOT is tightening (H1, confirmed). Neocloud H200 is absorbing price-sensitive long-context workloads below hyperscaler H100 rates (H2, confirmed). H100 ON_DEMAND sits between these two forces: above it, H200 neoclouds are cheaper; below it, L40S handles the inference workload that doesn't need 80 GB of HBM. Hyperscalers are administratively holding H100 ON_DEMAND steady as a strategic anchor price, but the market around it is repricing. When the freeze breaks — upward if the L40S tightening propagates into H100 demand, downward if neocloud H200 adoption accelerates — it will reprice every other tier simultaneously. The asymmetry favors watching for the upward break first, because the H100 is still the primary training and long-form inference chip, and L40S cannot fully substitute at scale.

Key Evidence:

1. AWS H100 price_changes pipeline: zero rows with >5% change over 7 days — structural price freeze confirmed, not missing data (price_changes API, Oct 2, 2026)
2. UPCOMING_PRICE_CHANGE alerts: zero across AWS, GCP, Azure, Oracle — no provider has telegraphed any near-term move (price_alerts API, Oct 2, 2026)
3. AWS Seoul H100 implied spot: $1.87/hr → $1.98/hr (+5.7% in 24h); spot discount compressed 75.2% (from 76.5%) — early tightening signal, LOW-MED confidence tier
4. H100 ON_DEMAND deltas: 0.00% at 24h, 7d, 30d in every region from Virginia ($8.69/hr) to Singapore ($3.35/hr) — 6 regions checked, all frozen (ticker data, Oct 2, 2026)
5. GCP Amsterdam H100 ON_DEMAND: fell from $0.336/hr to $0.146/hr (−56.7% in 24h per anomaly diagnostics) — identified as GCP DWS/committed-use rate leak, not a genuine list-price move; validates that the freeze is real in standard ON_DEMAND
6. Singapore H100 ON_DEMAND: +6.1% over 30d to $3.35/hr — the singular exception, and it is attributed to regulated supply caps, not organic demand (Singapore tracker confirms vertical supply curve)

Implied Action:

• Reserved instance arbitrage (moderate urgency): H100 RESERVED_1YR locks today's list price before any potential upward break. The risk/reward is asymmetric: downside is locking in a price that falls (unlikely near-term given L40S tightening dynamics); upside is locking the price before the first upward move, which would immediately ratchet ON_DEMAND rates. Monitor H100|RESERVED_1YR|us-east-virginia for rate changes as the earliest structured signal.
• Primary watch signal: First PRICE_CHANGE or UPCOMING_PRICE_CHANGE alert with severity ≥ MEDIUM on any H100 SKU from any provider. None exist today. That absence is informative but not permanent.
• Downgrade from "imminent": The 14–60 day framing was too aggressive. Revised to 30–90 days. The conditions are building, not yet triggering.

H4 — Singapore H100 Premium Validates the Regulated-Geography Moat Thesis — The Scarcest Compute Asset Is Approved Capacity

Confidence: 4 / 5

Thesis: Singapore is the only H100 ON_DEMAND market globally showing a positive 30-day price drift (+6.1% to $3.35/hr), and it is doing so while being 61% cheaper than the US-Virginia benchmark ($8.69/hr). This paradox — cheaper but rising, while more expensive markets are flat — is explained by a vertical supply curve enforced by government fiat. The Singapore datacenter tracker confirms: STT GDC (KKR/Singtel-backed, 2.3 GW of existing capacity, $10.9B investment) and Keppel DC Singapore (Keppel DC REIT, consolidating full ownership of DC3 and DC4 in 1H 2026) are the named operators of a formally allocated 200 MW of new capacity split among four licensed providers. Zero approved-stage Singapore datacenter projects appear in our pipeline — supply cannot be added without a new government allocation. The supply curve is literally vertical.

The Keppel DC REIT ownership consolidation of DC3 and DC4 is a specific, datable signal: operators are tightening control of existing capacity rather than waiting for new supply. When existing operators maximize utilization of allocated capacity while new supply requires regulatory approval, pricing power is structurally embedded. The 6.1% 30-day rise in Singapore H100 is not explained by demand alone — it is the first measurable expression of that pricing power in the rental market.

The broader thesis — that regulatory approval is the scarcest compute input — is supported by the construction pipeline. Of the 166 UNDER_CONSTRUCTION projects globally, the ones with the most certain revenue trajectories are those with named tenants, locked financing, and utility agreements (Amazon New Carlisle at 2.25 GW, $11B; Monarch Campus WV at 2 GW, $71B, Anthropic tenant; The Barn Michigan at 1.4 GW, Oracle/OpenAI). Project Jupiter/Stargate New Mexico at 4 GW — theoretically the largest — remains on hold due to permit denials and force majeure on its Bloom Energy fuel cell system, with project debt reportedly trading below 90 cents. The comparison is instructive: permitted and operating > funded and under construction > announced but unentitled.

Key Evidence:

1. Singapore H100 ON_DEMAND: $3.35/hr (+6.1% over 30d) — only positive 30d delta in ON_DEMAND H100 globally (ticker data, Oct 2, 2026)
2. STT GDC Singapore: 2.3 GW existing, $10.9B investment, KKR/Singtel ownership (datacenter tracker, confidence HIGH)
3. Keppel DC Singapore: Full ownership consolidation of DC3 and DC4 completed 1H 2026, driving record DPU growth (datacenter tracker)
4. Zero approved-stage Singapore datacenter projects in tracker — supply curve is vertical by government allocation (tracker query, Oct 2, 2026)
5. Singapore H200 ON_DEMAND: $13.10/hr — within 0.6% of Frankfurt H200 ($13.03/hr) despite Frankfurt's far larger supply base; Singapore premium is extending across generations (ticker data)
6. Project Jupiter/Stargate NM: ~26% complete, on hold, permit denial on Bloom Energy fuel cell system, debt below 90 cents — the cost of regulatory failure (datacenter tracker)
7. Singapore Jurong Island: 700 MW low-carbon allocation formally split among STT GDC, Digital Realty, Keppel, Equinix — government-issued oligopoly license (tracker)

Implied Action:

• Long Singapore-domiciled datacenter operators: Keppel DC REIT and STT GDC (private, KKR/Singtel) are the most direct expressions of the regulated-geography moat. These are not speculative buildout bets — they hold operational, allocated capacity. The record DPU from Keppel's DC3/DC4 consolidation is already pricing this in partially, but the H100 rental price drift (+6.1%) suggests the rental market is only beginning to reflect it.
• Long Equinix (EQIX) and Digital Realty (DLR) as the most liquid public vehicles with Singapore allocation presence. Any government announcement expanding the Jurong Island allocation — and naming these four operators — is a direct positive catalyst.
• Watch for Singapore H100 breaking $3.50/hr (currently $3.35/hr). A breach of that level on ON_DEMAND would confirm that regulated-market pricing power is compounding.
• Distinguish from the Warsaw artifact: Warsaw's $12.17/hr H100 ticker (n=1 provider, single decomposition) is not a regulated premium — it is a model artifact from an unusual CPU/RAM instance mix over-attributed to the GPU. Singapore's premium is real; Warsaw's is not.

Inventory & Channel Health

The inventory health chain is partially degraded today — the composite DIO regime and per-ticker inventory health tables returned empty, removing the distributor-lead signal from the cross-check. This is a data availability gap, not a market signal, and reduces confidence in any inventory-cycle claim this briefing would otherwise make.

What the trade flow data does confirm is directionally consistent with the demand-acceleration narrative: US flash/SSD imports (HS 852351) are running at $7.1B for July 2026, +499% YoY and +75% over 3 months (RISING signal). This aligns with AI data center storage buildout — NAND for KV-cache offloading, checkpoint storage, and inference serving infrastructure. However, the broad IC import flow (HS 8542, all chips) shows only +40% YoY with −13% MoM (FLAT signal), creating a structural mismatch: storage is being deployed faster than compute chips are arriving. The most probable interpretation is that data center construction is proceeding ahead of GPU rack delivery — buildings and storage are going in first, with GPU hardware following on the longer-lead supply chain.

One critical implication for the HBM scarcity narrative: with the distributor DIO chain unavailable, we cannot independently confirm whether Micron's "75%+ of 2027 HBM output pre-committed" claim (CEO statement, September 30) is reflected in channel inventory conditions. That claim remains unverified from our data. Treat HBM scarcity as a forward-looking capex risk corroborated by direction (flash imports surging, AMD stock +30% in 1 month) but not yet confirmed by current-period channel DIO reads.

Supply & Demand Pipeline

Demand (model releases — confirmed, high impact): The week of September 22–30 produced the densest frontier model GA burst this dataset has recorded: Grok 4.7 (xAI, 2.1T parameters, September 22, $2/$6 MTok), Claude Opus 5.5 and Sonnet 5.5 (Anthropic, September 28, $4/$20 and $2/$10 MTok respectively), Meta Muse enterprise agentic platform (September 28), and Gemini 4 Argon (Google, September 30, $1.47/$10 MTok per tracker, 1M output tokens). The causal timestamp on the L40S London spike ($0.73/hr in 48 hours coinciding with the September 28 launches) makes the model-release-to-GPU-demand transmission mechanism the most directly evidenced finding in this briefing.

Token price compression is real: Gemini 4 Argon's introductory $1.47/MTok input undercuts Claude Opus 5.5 at $4/MTok by 63%. This does not reduce GPU demand — it increases volume at lower margin — but it does suggest hyperscalers are absorbing inference margin as customer acquisition cost rather than passing it into rental prices. The inference efficiency gains (1M output token context, caching at 95% discount for Argon) reduce GPU-hours-per-token, partially offsetting volume growth. Net effect on GPU demand is positive but not proportional to raw token volume.

Supply (datacenter — capacity timing is the critical variable): The two most strategically significant upcoming capacity events are: (1) Monarch Compute Campus, Mason County WV (2 GW planned, 460 MW Phase 1, Anthropic tenant, NVIDIA Vera Rubin/Blackwell, certified off-grid microgrid) — targeting late 2027, designed explicitly to bypass the grid connection delays that are holding up Project Jupiter; and (2) The Barn Stargate Michigan (1.4 GW, Oracle/OpenAI tenant, first building nearly complete, targeting early 2028 operations, 19-year DTE Energy contract). Both are named-tenant, named-accelerator projects, meaning the GPU demand they represent is contractually committed, not speculative. Both target late 2027/early 2028 — which is precisely when HBM4/HBM4E volume production is expected to ramp and when NVIDIA's Vera Rubin/Blackwell Ultra generation should reach volume deployment. The 18-month gap to that supply event is the primary reason today's L40S/H100 pricing environment will persist longer than consensus may assume.

Project Jupiter/Stargate New Mexico (4 GW planned, ~26% complete) remains on hold due to permit denial on the Bloom Energy fuel cell system and Natural Gas pipeline delays, with project debt reportedly trading below 90 cents. If this resolves, it becomes the single largest capacity addition in the pipeline. If it does not, the 4 GW evaporates from the 2027 supply forecast entirely. This is the highest-stakes binary in the datacenter tracker.

Component supply: HBM4E from both SK Hynix and Samsung remains in SAMPLING status in the component tracker — not yet in volume production. This is consistent with the Monarch/Barn 2027–2028 deployment timeline but means the rental market operates on existing HBM3/HBM3e supply through 2026 and into 2027. Apple's AFM 3 Cloud Pro running on NVIDIA Blackwell (B200/GB200) in Google Cloud — a hyperscaler customer choosing NVIDIA GPU over Google's own TPU V6E for flagship production inference — is the clearest single data point confirming GPU remains the dominant inference substrate even within TPU-native infrastructure.

Pre-Announced Price Changes

None. Zero UPCOMING_PRICE_CHANGE alerts exist across AWS, GCP, Azure, and Oracle as of October 2, 2026. Every provider's published forward pricing pipeline is silent. Given the H100 ON_DEMAND freeze confirmed in the price_changes data, and the absence of any RESERVED rate adjustments in the pipeline, the most likely near-term pricing events remain in the SPOT tier (L40S EMEA, Inferentia Oregon, TPU V5/V6) and in neocloud markets that the alert pipeline does not cover (CoreWeave, Nebius, Hyperstack). The absence of any UPCOMING alert is itself an informative signal: no provider is telegraphing a price war, a discount event, or a rate increase for the near term.

Risk Flags

Immediate (0–30 days)